Normalization scheme to use for upstream sequences read count in (Chip-seq histones)

Question

[Saad Khan]

Hi,

I am comparing 1000bp upstream sequences of TSS amongst three species (without replicates). In order to compare the upstream 1000bp with the KA/Ks of the gene (pairwise comparison between species) I had normalized the 1000 bp upstream region using the rpkm formula.

I was wondering if this is the best normalization measure I can use in the absence of any replicates or are there other normalization methods I should try and what are they?

Regards

Answer 1

[Devon Ryan]

FYI, for fixed-width regions, RPKM devolves to CPM.

My main concern for something like this would be GC bias differences in the samples. Did you do input samples for each species as well? That'd allow you to correct the GC bias (see computeGCBias and correctGCBias in deepTools).

Aside from that, I would prefer to get scaling factors from non-peak regions, or at the very least to use a more robust method (TMM/RLE/quantile normalization). I think that any of these will be preferable to RPKM/CPM, since if you have any peaks with crazy amounts of duplicates (or that were simply absurdly highly covered) then they'll skew the stats.

[Saad Khan]

How do you get scaling factors from non-peak regions can you please elaborate. Also TMM/RLE and quantile normalization require replicates AFAIK.

PS :I did not have chip input samples

[Devon Ryan]

Replicates aren't needed or even used in TMM, RLE, or quantile normalization. All you need are multiple samples, which can be the ones being compared.

To get scaling factors from non-peak regions, you would first call peaks and then use the counts outside of them for the normalization.

Since you lack inputs, you're going to have more work on your hands when it comes to validation.

[Saad Khan]

So in my case I don't have multiple samples in one species I have different histone modifications in 3 species (for a particular tissue) that I am trying to compare here.

[Saad Khan]

Hi Devon,

Do you have something to elaborate on my question.

thanks

[Devon Ryan]

Not unless you'd like me to elaborate on something in particular.

[Saad Khan]

So I have only one sample for each species. What in that case would be the most correct way to normalize

CPM or TMM

[Saad Khan]

Is it possible to use TMM on such data and how?

[Devon Ryan]

Sure:

1. Call peaks
2. Make a bed file of some non-peak regions. Ideally there would be a good number of these (i.e., 1000 or so).
3. TMM
4. Scale or weight accordingly

Alternatively, play around with scale factors until things look right.

[Saad Khan]

So what you are saying that I should use the peak regions as sample1 and non-peak regions as sample2 for doing TMM normalization?

[Devon Ryan]

No, you would use the non-peak regions for all of the samples. TMM would give you scaling factors accordingly that you would then need to apply to the peak counts. This is the same as how ERCC spike-ins are used in RNAseq.

[Saad Khan]

Sorry I am still not able to follow.I have only one sample for each histone modification in one species. So if you are saying I need to use non-peak regions for all of the samples do you mean I should combine non-peak regions for all different histone modifications? Can you point me to an example?

[Devon Ryan]

Ah, right, I'd forgotten the context of your question. You'll just have to play around with TMM a bit if you want to use it. There are no examples of this that I'm aware of, but the general idea would be to ignore the fact that your non-peak regions are in different areas for each sample and to just use counts in some fixed number of them per sample. I don't have the time to put together a long example of this, unfortunately.